Learning Deconvolutional Network for Object Tracking

Object tracking can be tackled by learning a model of tracking the target's appearance sequentially. Therefore, robust appearance representation is a critical step in visual tracking. Recently, deep convolution network has demonstrated remarkable ability in visual tracking via leveraging robust high-level features. To obtain these high-level features, convolution and pooling operations are executed alternatively in deep convolution network. However, these operations lead to low spatial resolution feature maps which degrade the localization precision in tracking. While low level features have sufficient spatial resolution, their representation ability is insufficient. To mitigate this issue, we exploited deconvolution network in visual tracking. This deconvolution network works as a learnable upsampling layer which takes low-resolution high-level feature maps as input and outputs enlarged feature maps. Meanwhile, the low level feature maps are fused with these high level feature maps via a summarization operation to better represent target appearance. We formulate the network training as a regression issue and train this network end to end. Extensive experiments on two tracking benchmarks demonstrate the effectiveness of our method.